Link - Irisa

Locality-aware and lazy overlay networks for WANs
About Inria
Inria, the French national institute for research in computer science and control, is
dedicated to fundamental and applied research in information and communication
science and technology (ICST). Inria has a workforce of 3,800 people working
throughout its eight research centers established in seven regions of France.
The PhD student recruited will be directed by Christine Morin from the Myriads team
and co-advised by Marin Bertier from the ASAP team and C dric Tedeschi from the
Myriads research team, located at the Inria Rennes center.
Mission and activities
This PhD position targets the design, development and experimental validation of the
overlay network envisioned in the Discovery initiative [1]. 1 This next-generation
overlay aims at offering the best trade-off between maintenance cost and efficiency,
while being built according to physical locality between nodes.
While the Internet was built in a decentralized fashion, we are currently witnessing a
strong re-centralization of computing facilities such as those run by Amazon, Google
or Facebook. The intrinsic limitations to this model is increasingly visible, chiefly due
to reliability, privacy, and energy issues, as is experienced recently by users [5].
The DISCOVERY initiative [2] is driven by the objective of proposing a decentralized
cloud architecture composed by leveraging on the resources within the network’s
backbone. Telecom operator’s infrastructure appears to be highly over-provisioned.
In other terms, the backbone links and routers are highly underutilized most of the
time. This suggests that this infrastructure could be exploited to underlie a utility
computing platform. By adding computing resources on Point-of-presence (PoPs), a
distributed utility computing platform could be deployed inside the telecom operator’s
core network. Doing so, a significant economy of scale can be achieved, with cost
savings related to the construction, maintenance and energy supply of such a
computing center, by capitalizing on the existing infrastructure. Also, it will lead to a
natural distribution of resources while favoring the use of resources that are
geographically close to the user.
Exploiting such a utility computing platforms calls for mechanisms to place
applications, under the shape of virtual environments (made of multiple cooperating
virtual machines), efficiently. The different problems to address are the following.
Firstly, the need is to locate close nodes. Indeed, deploying a set of cooperating
virtual machines on a set of physically close nodes will avoid increasing latencies
and uselessly burdening the network. Secondly, we need to provide a data
management layer allowing to retrieve data efficiently while minimizing transfers
when concurrent geographically-distributed accesses are made on these data. For
instance, creating a data can always lead to a local storage of this data. Then, data
migrations due to subsequent read and write operations can be minimized, by using
indirection mechanisms for subsequent data retrieval operations. Thirdly, ensuring
data availability will require devising a replication scheme. The replication
mechanisms designed must, again, account for a minimization of the data transfers
The work in this thesis will consist in proposing the building blocks for the design and
implementation of reliable service and data placement mechanisms to be used in the
DISCOVERY initiative. On the conceptual side, this will call for building proximityaware overlay networks [3], large data (key value stores) management [7] and lazy
data maintenance schemes [6].
The work will go through several phases. Firstly, the requirements will have to be
formalized. Secondly, the algorithms will be developed and validated through
analysis and simulation. Thirdly, an experimental validation and tuning of the
parameters of the algorithm, probably through their deployment on a large-scale
platform such as Grid'5000 [4] will be planned.
[1] M. Bertier, F. Desprez, G. Fedak, A. Lebre, A.-C. Orgerie, J. Pastor, F. Quesnel, J.
Rouzaud- Cornabas, , and C. Tedeschi. Beyond the clouds: How should next generation
utility computing infrastructures be designed? In Z. Mahmood, editor, Cloud Computing:
Challenges, Limitations and R&D Solutions. Springer, Computer Communications and
Networks, Springer, 2014.
[3] Frank Dabek, Russ Cox, M. Frans Kaashoek, Robert Morris: Vivaldi: a decentralized
network coordinate system. SIGCOMM 2004: 15-26
[6] Sergey Legtchenko, S bastien Monnet, Pierre Sens, Gilles Muller: RelaxDHT: A churnresilient replication strategy for peer-to-peer distributed hash-tables. TAAS 7(2): 28 (2012)
[7] Cassandra: Principles and Application. Dietrich Featherston. Department of Computer
Science University of Illinois at Urbana-Champaign.
Skills and profiles
Knowledge and/or experience in distributed systems
Experimentation skills (simulation and in-vivo experiments)
Highly motivated, autonomous and curious
English mandatory (French is a plus)
Additional information
The candidates are invited to contact Marin Bertier and C dric Tedeschi
([email protected]).
Duration: 36 months
Location: Rennes, France
Monthly salary before taxes: 1958 euros (medical insurance included).